Preserving Product Geometry in Generative E-Commerce Pipelines: Behind Propulsell

The Fidelity Problem in Generative Staging

Early diffusion models struggled with commercial e-commerce applications because of hallucination: logos distorted, stitch patterns warped, and garment dimensions drifted from physical merchandise reality.

For a catalog, those errors are not cosmetic. A product image is a promise about what the customer will receive. If a label is misspelled, a strap changes length or a fabric gains a texture it does not have, the image misrepresents the item. Marketplaces generally require listing images to accurately depict the product being sold, and shoppers who receive something that looks different from the photo are more likely to return it.

Propulsell solves this through a multi-stage visual synthesis pipeline that decouples the product subject from the generated environment.

Decoupling the Product from the Scene

The core idea is simple to state: the model should generate the world around the product, not the product itself. If the pixels that depict the merchandise are never regenerated, there is nothing for the model to distort.

In practice this means treating every image as two layers. The subject layer comes from the original photograph and carries the real shape, color, texture and typography. The scene layer is generated: a surface, a backdrop, props and lighting that fit the brand's art direction. The difficulty, and most of the engineering, lies in making those two layers look like they were captured by the same camera at the same moment.

Edge-Constrained Masking and Depth Conditioning

Before any background generation occurs, high-precision boundary analysis isolates the source product. Depth estimation matrices define exact spatial coordinates, ensuring that generated surfaces (such as marble tables, wooden desks, or outdoor paving) realistically interact with the product's physical boundaries.

Edges are where composites usually fail. Hair-thin straps, translucent bottles, fur and mesh all produce boundaries that are neither fully product nor fully background. A mask that is too tight clips the product; one that is too loose drags the old background along with it. Constraining generation to respect a precise boundary keeps the silhouette intact.

Depth matters because objects rest on surfaces. A product that is pasted onto a scene without a believable contact point appears to float. Using depth information, the scene can be generated so that a tabletop meets the base of a bottle at the right height and perspective, and so that soft contact shadows appear where the product touches the surface.

A secondary conditioning pass recalculates directional light vectors so highlights across the subject accurately match ambient illuminance in the synthesized scene.

Matching Light and Shadow

Viewers notice inconsistent lighting even when they cannot say what is wrong. If the scene suggests a window on the left but the product's highlights fall on the right, the image reads as fake. Matching the direction, softness and color temperature of the light across both layers is what makes a composite feel photographed rather than assembled.

Shadows carry the same weight. A contact shadow anchors the product to the surface, and a cast shadow consistent with the scene's light source places it in the room. Getting both right is often the difference between an image that converts and one that quietly erodes trust in the listing.

Verification Before Publishing

Generation is only half of the pipeline. Like every ADERYX product, Propulsell treats generated output as untrusted until it has passed validation. Checks run before an image is persisted, so a render that drifts from the source product is caught in the pipeline instead of on a live product page.

This is the same principle we apply across our portfolio: AI produces candidates, and the product decides what is good enough to ship. It is what lets a catalog team rely on batch output without reviewing every pixel by hand.

Preparing Source Photos for the Best Results

Better inputs produce better outputs, whatever tool you use. Photograph the full product in frame with a little margin around it, at the highest resolution you have. Use even, diffuse light so that strong shadows and reflections from the original setup do not fight the new scene.

A plain, contrasting background makes the product boundary easier to isolate. Keep labels facing the camera and in focus if typography matters, and shoot each angle you want to sell separately rather than relying on generation to invent a view you did not capture.

Production Scalability Across SKU Catalogs

For enterprise brands launching hundreds of SKUs per season, manual re-touching is infeasible. Propulsell orchestrates asynchronous cloud GPU workers to batch render standardized lighting templates, outputting web-optimized WebP assets directly into e-commerce CMS systems.

Templates are what keep a large catalog coherent. When every product in a collection is rendered with the same surface, camera height and lighting, the category page looks art-directed instead of assembled from different shoots. Running that work asynchronously means a catalog upload does not block the team; renders arrive as they finish and can be reviewed in batches.

Key Takeaways

Generative staging is only useful for commerce if it never alters the product. Propulsell keeps the original product pixels, generates only the scene, and uses precise masking, depth and lighting conditioning to make the two layers agree. Validation before publishing and template-based batch rendering make that approach practical across entire catalogs.